Reproduce “Self-supervised DXA representations encode multi-system disease risk, biological aging and heritability” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gil Sasson, Zachary Levine, Smadar Shilo, Sarah Kohn, Guy Lutsker, Anastasia Godneva, Adam Gabet, David Krongauz and 4 more
LeDXA learns a representation of spatial information in unlabeled DXA scans and uses it for health prediction. It learns from scan views, not disease labels, and was tested in an independent UK Biobank cohort. Over median 4.3-year follow-up, it outperformed scanner-derived DXA measurements for 9 of 20 incident endpoints, including hip arthrosis. These observational results do not establish causes or improved outcomes.
Key insights
- Raw DXA images contain rich, underutilized spatial and textural information about systemic health that surpasses the diagnostic and prognostic value of conventional scalar measurements.
- Domain-specific self-supervised learning can effectively extract genetically meaningful and prognostically valuable representations from medical imaging data, outperforming general-purpose vision models and traditional metrics.
- LeDXA representations enable the derivation of a "biological-age gap" that correlates with mortality and disease burden, and facilitate the discovery of novel, physiologically distinct body composition phenotypes from unsupervised clustering.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.